W2C and extrinsics

[haowang020110]

Dear author
Sorry I have a name problem
It seemed that the normal conception of the camera extrinsics is w2c matrix.
However you made extrinsics=c2w, can you tell me the reason?
Sorry again for my poor knowledge.

mvsplat/src/dataset/dataset_re10k.py

Lines 224 to 226 in ef976a1

	w2c = repeat(torch.eye(4, dtype=torch.float32), "h w -> b h w", b=b).clone()
	w2c[:, :3] = rearrange(poses[:, 6:], "b (h w) -> b h w", h=3, w=4)
	return w2c.inverse(), intrinsics

[donydchen]

Hi, @haowang020110, there is no specific reason for using c2w or w2c. The main reason might be the encoder involves some operations that are easier to start from camera space, e.g., the wrapping operation (#80), and the encoder predicts 3D Gaussians in camera space, which is then transformed into world space. Hence, it seems more convenient to use c2w in this project.

[haowang020110]

Thanks a lot!